An interview recorded on two independent microphones is not the same data problem as a panel captured by one room microphone. Speechmatics distinguishes speaker diarization, channel diarization, and combined approaches; choosing the wrong one can misassign quotations even when every word is recognized correctly.
Three diarization routes and their differences
Speechmatics describes speaker diarization as grouping words by apparent voice identity. Channel diarization instead treats separately recorded audio channels as independent streams. The documentation also describes combined channel-plus-speaker diarization, available for realtime use, for cases where voices must be distinguished inside each audio channel. Speaker labels do not reveal legal identity, and stereo channels do not always correspond to people: a conference mixer may put both participants in both channels. Verify the recording topology before choosing an API setting.
Realtimes, batches and agent modes are not identical
The product documentation lists diarization across Batch, Realtime and Agent STT but notes that the Agent STT mode supports speaker diarization rather than channel diarization. An application designed for a telephone-agent turn-taking system should therefore not assume the same channel controls exist in prerecorded transcription. Also review model selection separately: current Speechmatics language documentation differentiates Enhanced and Standard families from multilingual Melia 1 and Oak 1, with different auto-language and translation behavior. Document both API mode and model identity in exported editorial records.
Overlap, stereo and identity traps
When two people speak at once, a diarizer may assign only one speaker or create a rapid sequence of switches. If a remote guest’s voice leaks into the host microphone, channel labels may no longer represent clean participants. Create a manual ground truth for at least a few overlapped segments and verify which timestamps and speaker IDs survive edits. For a legal or investigative quotation, a speaker ID requires human linking to a known person using independent context; an ID such as speaker_1 is not evidence that it belongs to the interviewee.
Selection protocol
Start by counting sources: one room microphone suggests speaker diarization; clean isolated guest tracks suggest channel diarization; mixed isolated tracks may need combined handling if a particular endpoint supports it. Test the same clip with and without diarization, comparing speaker-attributed word error and the manual correction time. Store separate field names for `channel`, `speaker_label`, and `verified_person`; never overwrite a raw model label with a human name without a review log. Follow the provider’s current API examples for accepted parameter names rather than copying a configuration from an earlier version.
Publication and data governance
Meeting participants may consent to a private transcription but not to publication of a speaker-attributed public archive. Before transmitting audio, obtain required workplace or interview permissions, assess processing regions and retention settings, and define which speakers may be named. In sensitive settings, do not use automated diarization as authentication. If the recording is primarily intended for captions, a descriptive label such as ‘Host’ or ‘Audience member’ may communicate more accurately than a guessed personal name.
Worked production example
Consider a remote podcast: the host is on left-channel mic one, the guest on right-channel mic two, and a group discussion later comes through a single mixed call. Keep the isolated interviews as channel-separated data while using voice-based diarization for the mixed recording. Record the exact hardware changes: in the mixed section, a channel label can no longer be assumed to identify a single person. Compare overlap segments manually before converting speaker labels to named captions.
Practical acceptance gate
Require a channel-layout diagram, source-audio hash, sample speaker-identity review, overlap error analysis and a named-person signoff for any sensitive quotation. Test the chosen Batch or Realtime mode, because combined options are not globally identical. If the application runs Agent STT, block channel-diarization features that the provider does not document for that mode.
Region, plan and model-version notes
Documentation checked on 2026-10-25. Exact applicability: availability varies by account/endpoint and jurisdiction. Version: Batch / Realtime / Agent STT (per endpoint). Plan: not independently verified for a specific plan. Monetary prices and quotas are not asserted unless separately verified; check the canonical AI 10 pricing ledger before purchase or publication of a tariff.
Sources & Further Reading
- Speechmatics Diar — official documentation (checked 2026-10-25)
- Speechmatics Languages — official documentation (checked 2026-10-25)